Fantastic linear systems and how to solve them - Pleasingly parallel computation of resolution and uncertainty in seismic tomography
<p>In classic global seismic tomography one ultimately needs to solve a large, typically sparse, linear system. Specific formulations of the tomographic problem then allow for computing the resulting model, as well as resolution and uncertainty. Although the latter two properties are crucial for the assessment of a given tomographic image, they are often not calculated because of high computational costs.</p> <p>Based on an actual data set used for global tomography, we discuss the usual and an alternative tomographic system and investigate the possibility of embarrassingly parallel (or pleasingly, perfectly parallel) computations of the full model result including resolution and uncertainty. The necessary numerical strategies and computational concepts are briefly explained from a practical point of view. This includes possible iterative and direct numerical solvers like, for example, LSQR and QR decomposition, a discussion of multiprocessing vs. multithreading and first personal experience with GPU computing using the programming language Julia.</p>
https://www.geophysik.uni-muenchen.de/en/seminars/seminars/geocomputing-29/t-b-a-6
https://www.geophysik.uni-muenchen.de/@@site-logo/kopfbildvorlage_Geophysik_23_03_22.jpg
Fantastic linear systems and how to solve them - Pleasingly parallel computation of resolution and uncertainty in seismic tomography
Abstract
In classic global seismic tomography one ultimately needs to solve a large,
typically sparse, linear system. Specific formulations of the tomographic
problem then allow for computing the resulting model, as well as resolution
and uncertainty. Although the latter two properties are crucial for the
assessment of a given tomographic image, they are often not calculated
because of high computational costs.
Based on an actual data set used for global tomography, we discuss the usual
and an alternative tomographic system and investigate the possibility of
embarrassingly parallel (or pleasingly, perfectly parallel) computations of
the full model result including resolution and uncertainty. The necessary
numerical strategies and computational concepts are briefly explained from a
practical point of view. This includes possible iterative and direct
numerical solvers like, for example, LSQR and QR decomposition, a discussion
of multiprocessing vs. multithreading and first personal experience with GPU
computing using the programming language Julia.